A Fine-Grained Latent Aspects Model for Recommendation: Combining Each Rating with Its Associated Review

Abstract

Recently, several approaches simultaneously exploiting ratings and review texts have been proposed for personalized recommendations. These approaches apply topic modeling techniques on review texts to mining major latent aspects of the item (or the user) and align them with collaborative filtering algorithms to increase the accuracy and interpretability of rating prediction. However, they learn the topics for each item (or user) by harnessing all reviews related to it, which is not intuitive or in line with users’ rating and review behavior. In this paper, we propose a Fine-grained Latent Aspects Model (FLAM), which learns the topics for each review with the corresponding latent aspect ratings of the user and the item. FLAM is an united model of Latent Factor Model (LFM) and Latent Dirichlet Allocation (LDA). LFM, well-known for its high prediction accuracy, is employed to predict latent aspect ratings of the user and the item. LDA, a classical topic model, is used to extract latent aspects in the reviews. Our experiment results on 25 real-world datasets show the proposed model has superiority over state-of-the-art methods and can learn the latent topics that are interpretable. Furthermore, our model can alleviate the cold-start problem.